[20240712_LabSeminar_Huy]Spatio-Temporal Neural Structural Causal Models for Bike Flow Prediction.pptx

Spatio-Temporal Neural Structural Causal Models for Bike Flow Prediction

1. Quang-Huy Tran Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: [email protected] 2024-07-12 Spatio -Temporal Neural Structural Causal Models for Bike Flow Prediction Pan Deng et al. AAAI- 202 3 : The Thirty-Seventh AAAI Conference on Artificial Intelligence

2. OUTLINE MOTIVATION METHODOLOGY EXPERIMENT & RESULT CONCLUSION

3. MOTIVATION Bike-Sharing systems have been widely deployed in urban public tran sportation: Convenience and environmental friendliness in recent years . E ffective allocation of bike-sharing resources to enhance the quality of system service is important . However, due to high frequency and randomness of bike usage throughout city, stations often get imbalanced over time . System congestion and insufficient stations . Necessary to accurately predict the bike flow in each commuting area. Bike usage patterns and spatio -temporal dependencies between regions are affected by external conditions. many confounding factors and extract spurious correlations are showed up . Previous works were not considering. Overview and Challenges

4. INTRODUCTION Propose a counterfactual representation reasoning module : extrapolate the spatio -temporal state under the factual scenario to the future counterfactual scenario. enhances the feature’s understanding of future states. Contribution Provide a novel causality-based interpretation for the bike flow prediction : apply the front door criterion based on causal interventions to remove confounding biases in the feature extraction process .

5. METHODOLOGY Structural Causal Model Formulate Causalities among bike flow , contextual condition . can decouple into spatial neighborhood features and temporal dynamic features . and integrate into spatio -temporal states .   Problem: predicted target .   Graph construction Geographical distance graph and transition probability graph based on the historical trip records of the regions.   where is distance between region and calculated by the latitude and longitude of regional center, is distance threshold. is variance of distance matrix and is transition flow.  

6. METHODOLOGY Main Architecture

7. METHODOLOGY Causal Intervention via Frontdoor Criterion A pply the front door criterion based on path Cutting off link .   Propose Input Gate to fit prior distribution, Dynamic Causality Generator to embed spatio -temporal causality into a dynamic causal graph, and Spatio -Temporal Evolutionary Graph Convolution to extract spatio -temporal states. generation of time-varying spatio -temporal states from features to describe the spatiotemporal patterns inherent in the data prior distribution of the input data . process of extracting spatio-temporal features from data and noise.

8. METHODOLOGY Causal Intervention via Frontdoor Criterion Input gate: Take the historical periodic flow data as input: previous week, previous day, and previous time steps. Fully Connected Layer and Concatenation of the periodic flow data and the external conditions within the same time step to obtain . Gated linear unit is used to output the context conditioned features   where are model parameters, is the element-wise product, is tanh function, and is sigmoid function  

9. METHODOLOGY Causal Intervention via Frontdoor Criterion Dynamic Causality Generator : Inter-regional relationships are affected not only by spatio -temporal causality, but also by diverse external conditions. Coupled spatio -temporal states and context-conditioned features . Input of dynamic causality generator with squeeze excitation method: A self-gating mechanism based on node dependence. where are model parameters and d is the number of feature channels .   C alculate inter-node similarity and embed the dynamic causality into causal graph: where are model parameters. is sigmoid function .  

10. METHODOLOGY Problem Definition Spatio -Temporal Evolutionary Graph Convolution: Simplify and adopt diffusion convolution to s eparately propagate the inflow and outflow information of each node where n is depth of p ropagation, are model parameters; are contribution coefficient . Normalized adjacency matrix:  

11. METHODOLOGY Causal Intervention via Frontdoor Criterion where are the parameters of graph convolution and is the spatio -temporal state of the STNSCU at time step t.   Spatio-Temporal Neural Structural Causal Unit : Integrate first-three t o represent the complete causal intervention process in form of encoder-decoder structures.  

12. METHODOLOGY Counterfactual Representation Reasoning Under the condition of in the factual scenario, to predict C is set to to infer the spatio -temporal state in the counterfactual scenario.   Where represents counterfactual representation reasoning process, and is prediction process .   where d is number of feature channels. are learnable parameter .   T o infer spatio -temporal state in case of by focusing on the similar part between external conditions of future and history.   Future representations are input into a FC layer and then used to initialize the decoder.    

13. EXPERIMENT AND RESULT EXPERIMENT SETTINGs Dataset: NYC- Bike and BJ- Bike . Baselines: Spatial-Temporal GNN: STGCN[1], STGODE[2], CCRNN[3], DMSTGCN[4], GMAN[5], ASTGNN[6], and DGCRN[7]. Measurement : Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error(MAPE) . [1] Yu, B., Yin, H., & Zhu, Z. (2017). Spatio -temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875. [2] Fang, Z., Long, Q., Song, G., & Xie, K. (2021, August). Spatial-temporal graph ode networks for traffic flow forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining (pp. 364-373). [3] Ye, J., Sun, L., Du, B., Fu, Y., & Xiong, H. (2021, May). Coupled layer-wise graph convolution for transportation demand prediction. In Proceedings of the AAAI conference on artificial intelligence (Vol. 35, No. 5, pp. 4617-4625). [4] Han, L., Du, B., Sun, L., Fu, Y., Lv , Y., & Xiong, H. (2021, August). Dynamic and multi-faceted spatio -temporal deep learning for traffic speed forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining (pp. 547-555). [5] Zheng, C., Fan, X., Wang, C., & Qi, J. (2020, April). Gman : A graph multi-attention network for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence (Vol. 34, No. 01, pp. 1234-1241). [6] Guo, S., Lin, Y., Wan, H., Li, X., & Cong, G. (2021). Learning dynamics and heterogeneity of spatial-temporal graph data for traffic forecasting. IEEE Transactions on Knowledge and Data Engineering, 34(11), 5415-5428. [7] Li, F., Feng, J., Yan, H., Jin, G., Yang, F., Sun, F., ... & Li, Y. (2023). Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution. ACM Transactions on Knowledge Discovery from Data, 17(1), 1-21.

14. EXPERIMENT AND RESULT RESULT – Overall Perfor mance

15. EXPERIMENT AND RESULT RESULT – Visualization The change process of the contribution coefficient during the training period   Static topologies and dynamic causal graphs

16. CONCLUSION B uild a causal graph to describe the traffic prediction problem from a perspective of causality due to the disturbance of incomplete observation, there are spurious correlations in the feature extraction process . resulting in the model can only perform general scenarios but failing in special scenarios. Summarization Propose a novel spatio -temporal neural structural causal model: D ecompose the front door criterion into multiple sub-terms . Main d ynamic causality generator module: e mbed the inter-regional time-varying causal relationship into the dynamic causal graph, enable to capture the dynamic rules . A counterfactual representation reasoning module: make spatio -temporal states in the current factual scenarios can represent count erfactuals.